OpenPipe's hosted platform stopped training and serving models on July 30, 2026, after CoreWeave acquired the company (announced September 3, 2025), so former users need a new home: W&B Training is the official successor, Together AI is the closest managed per-token replacement, Tinker and Unsloth suit teams who want more control, and Kojen fits if what you really need is cheaper models that pass your tests. OpenPipe's open-source RL library, ART, is still maintained.
OpenPipe was popular for distilling an expensive model's behaviour into a cheaper fine-tuned one, so the right replacement depends on whether you still need trained weights.
OpenPipe's training and inference moved to Weights & Biases and CoreWeave: serverless SFT and RL cost $2.70 per GPU-hour of active training, and tuned LoRA adapters serve on W&B Inference. ART remains free and Apache-2.0 for running RL on your own GPUs. Managed evals, reward models and DPO from the old platform were not carried over.
Pricing: ART: free, Apache-2.0. Successor CoreWeave serverless SFT and RL: $2.70 per GPU-hour of active training, with rollouts and storage billed separately. Source · Open source: yes (Apache-2.0 (ART))
Strengths
ART is one of the best open-source libraries for training agents with reinforcement learning.
RULER removes the need to hand-write reward functions.
Managed GPU-hour pricing on a large cloud through W&B Training.
Limits
The no-code fine-tuning SaaS, with its managed evals and DPO, no longer exists.
Best for: managed per-token fine-tuning of open models with downloadable weights
Together AI publishes per-token prices (SFT from $0.34 per million tokens for 8B-class models), supports LoRA and full fine-tuning with SFT and DPO, lets you download the weights, and offers LLM-judged evaluations and dedicated endpoints. Dedicated endpoints bill while running.
Pricing: SFT from $0.34 per 1M tokens for 8B-class models up to $2.03 for Llama 3.3 70B (DPO costs more; $4 minimum per job). Post-Training on B200: $11/GPU-hour. Source · Open source: no
Strengths
Transparent per-token prices on a wide range of open-weight models.
You can download the tuned weights or adapter.
Evaluations, fine-tuning, RL and serving on one platform.
Limits
Dedicated endpoints bill while running, even when idle.
RL is a separate GPU-minute product that needs credit for four hours up front.
Best for: researchers who want to write their own training loop without managing GPUs
Thinking Machines' Tinker runs your SFT, DPO or RL loop as LoRA on its GPUs across many models, priced per token (Qwen3-8B training at $0.44 per million tokens). You write the code, and production serving is limited.
Pricing: Per 1M tokens, by model: e.g. Qwen3-8B training $0.44; larger models cost more. Checkpoint storage $0.10/GB-month. Source · Open source: no
Strengths
Full control of the training algorithm without managing GPUs.
Unsloth's open-source library is the fastest, lowest-memory way to run LoRA, QLoRA, full fine-tuning, DPO or GRPO on a single GPU. You supply and manage the hardware and the serving.
Pricing: Open source: free. Pro and Enterprise tiers (multi-GPU, multi-node) are listed without public prices. Source · Open source: yes (Apache-2.0 core; some parts AGPL-3.0)
Strengths
The fastest, least memory-hungry way to fine-tune on one GPU, with a huge community.
Covers everything from QLoRA to full fine-tuning and GRPO reinforcement learning.
Free and runs entirely on hardware you control.
Limits
You supply, run and pay for the GPUs.
No evaluations, red-teaming or hosted deployment; paid-tier prices aren't published.
Best for: replacing a distilled model with a tuned prompt on a cheaper hosted model
If your OpenPipe fine-tune existed to make a cheaper model behave like an expensive one, Kojen's model search and prompt tuning may get you there without training: it tests 11 hosted models against your cases, reports accuracy with confidence intervals, and serves the winner behind an OpenAI-compatible endpoint. Text LoRA on 8B models is early access and can't be served yet. We make Kojen.
Pricing: Free to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source · Open source: no
Strengths
Measures and improves in one loop: evaluations, red-teaming, prompt tuning, model search and deployment share the same project and test cases.
Honest statistics by default: every rate has a confidence interval, tuned prompts must beat the current one in a paired test, and pass@k / pass^k come from several samples per case.
Agent-native: the full loop is exposed over an MCP server and a REST API with an OpenAPI spec, with the same 24 tools as the in-app agent.
Limits
Young product: weight training is early access on 8B models, LoRA adapters can't be served yet, and full fine-tuning and RLHF are previews.
No tracing SDK for apps running elsewhere, so it is not a replacement for an observability platform.
Closed source and hosted only; no self-hosting.
Side-by-side comparison
"Partial" means the feature exists with real limits, given in the note. "Not verified" means we couldn't confirm it from the vendor's own material.
OpenPipe alternatives: features and pricing, checked November 2026
Checked Nov 15, 2026
OpenPipe (now ART + W&B Training)
Together AI
Tinker
Unsloth
Kojen
What it is
Fine-tuning
CoreWeave (acquisition announced September 3, 2025)
Fine-tuning
Together AI (independent)
Fine-tuning
Thinking Machines Lab
Fine-tuning
Unsloth AI (independent)
Training dojo
Kojen (that's us)
Pricing
ART: free, Apache-2.0. Successor CoreWeave serverless SFT and RL: $2.70 per GPU-hour of active training, with rollouts and storage billed separately. Source
SFT from $0.34 per 1M tokens for 8B-class models up to $2.03 for Llama 3.3 70B (DPO costs more; $4 minimum per job). Post-Training on B200: $11/GPU-hour. Source
Per 1M tokens, by model: e.g. Qwen3-8B training $0.44; larger models cost more. Checkpoint storage $0.10/GB-month. Source
Open source: free. Pro and Enterprise tiers (multi-GPU, multi-node) are listed without public prices. Source
Free to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source
Run a model or app over a set of test cases and score the results.
No
The old platform's managed evals were not migrated
Yes
Classify, score and compare evaluations
No
No
Yes
Test cases scored for accuracy, hallucination, bias and latency
LLM-as-judge scoring
A model grades outputs against a rubric or reference.
Yes
RULER, an LLM judge that produces RL rewards
Yes
No
No
Yes
Judge kept out of the target's model family in simulations
Confidence intervals
Rates reported with an interval or a significance test, not just a point estimate.
No
Not verified
No
No
Yes
Wilson and bootstrap intervals on every rate
Prompt optimization
Automatically rewrites instructions or examples to raise a metric.
No
Not verified
No
No
Yes
Prompt tuning plus search across 11 hosted models
Official MCP server
An MCP server from the vendor that lets agents drive the product.
No
Has an example that trains models to use an MCP server's tools
Not verified
Not verified
No
Local models can be used from MCP clients; no Unsloth MCP server
Yes
ai.kojen/kojen in the MCP Registry, OAuth sign-in
Self-host
You can run it on your own infrastructure.
Yes
ART runs on your own GPU
Partial
Weights downloadable; the service itself is hosted
No
Yes
Runs on your own GPUs
No
Hosted only
What happened to OpenPipe
CoreWeave announced its acquisition of OpenPipe on September 3, 2025. OpenPipe then moved its training and inference to Weights & Biases, which CoreWeave also owns, and stopped training new models and serving existing ones on openpipe.ai on July 30, 2026. The site now carries a migration notice. The open-source Agent Reinforcement Trainer (ART) and its RULER judge for automatic rewards continue.
Two other names often suggested as alternatives have also changed: Predibase was acquired by Rubrik in June 2025 and its site now redirects to Rubrik Agent Cloud, with its LoRAX serving repo archived; and Fireworks AI has paused managed fine-tuning in favour of a code-first Training API.
When to choose OpenPipe (now ART + W&B Training), and when to choose Kojen
Choose OpenPipe (now ART + W&B Training) if…
You already use ART for agent RL and want the managed version on W&B Training.
You want to keep LoRA adapters served on the same cloud that trained them.
Choose Kojen if…
You want to check whether a tuned prompt on a cheaper model already passes your tests.
You want evaluations with confidence intervals and red-teaming around whatever you ship.
You want an OpenAI-compatible endpoint without running your own serving.
Frequently asked questions
Is OpenPipe shut down?
Its hosted platform is. OpenPipe stopped training and inference on openpipe.ai on July 30, 2026, after moving them to Weights & Biases and CoreWeave. The open-source ART library is still maintained.
Who owns OpenPipe?
CoreWeave, which announced the acquisition on September 3, 2025. CoreWeave also owns Weights & Biases, where OpenPipe's training moved.
What does W&B Training cost?
As of November 2026, CoreWeave lists serverless SFT and RL at $2.70 per GPU-hour of active training, with rollout generation, evaluation and checkpoint storage billed separately.
Is Predibase an OpenPipe alternative?
Not as a standalone product any more. Rubrik acquired Predibase in 2025, predibase.com redirects to Rubrik Agent Cloud, and its LoRAX repository was archived in October 2026.
Run an LLM-judged evaluation with confidence intervals, a red-team pass and a prompt-tuning run on your agent. Free to start, no card needed, then pay as you go.